Fix double exp() of logit_scale in create_probabilities

[Copilot]

BaseClassifier.create_probabilities() was calling self.model.logit_scale.exp(), but the BioCLIP pretrained model stores logit_scale already as the scale (not its log), so this applied exp() twice — producing inflated logits and over-confident (potentially overflowing) probabilities.

Changes

• src/bioclip/predict.py: Remove .exp() from logit_scale in create_probabilities():

  # Before
  logits = (self.model.logit_scale.exp() * img_features @ txt_features)
  # After
  logits = (self.model.logit_scale * img_features @ txt_features)

  Mirrors the identical fix in Imageomics/bioclip#39.

• tests/test_predict.py: Add TestCreateProbabilities with two unit tests — one asserting the corrected direct-scale behavior, and one confirming the double-exp() path produces a meaningfully different (wrong) result.

[hlapp]

@work4cs your change seems to lead to complete mis-prediction here (the test failure is in classifying the cat as a plant rather than a cat, not in the newly introduced unit test). Any idea why this is?